policy.md 26 KB
% Binary blob extermination policy % Leah Rowe % 2 January 2022 (updated 23 January 2022)
This article was written by Leah Rowe, the founder and current lead developer of Libreboot.
Introduction
Libreboot intentionally de-blobs coreboot, which is to say that it does not include binary blobs. The coreboot software otherwise requires binary blobs on most systems that it has support for. Libreboot's version of coreboot is entirely free, on its consequently reduced set of supported mainboards.
Libreboot is designed to comply with the Free Software Foundation's Respects Your Freedom criteria and the GNU Free System Distribution Guidelines (GNU FSDG), ensuring that it is entirely Free Software.
It was decided that a formal policy should be written, because there is quite a bit of nuance that would otherwise not be covered. Libreboot's policies in this regard were previously ill defined.
Background information
Libreboot concerns itself only with what goes in the main boot flash IC, but there are other pieces of firmware to take into consideration, as covered in the Libreboot FAQ.
Most critical of these are:
- Embedded controller firmware
- HDD/SSD firmware
- Intel Management Engine / AMD PSP firmware
Specific binary blobs are also problematic, on most coreboot systems, but they differ per machine. Libreboot excludes binary blobs in releases, so it only supports a handful of machines from coreboot.
For information about Intel Management Engine and AMD PSP, refer to the FAQ.
So what is Libreboot's policy?
Libreboot follows a very conservative and light touch approach, when it comes to deblobbing coreboot.
Libreboot only excludes software binary blobs, plus CPU microcode updates, completely in line with FSF policy. In practise, it is mostly microcode updates that Libreboot's build system deletes, along with coreboot Git history so that no traces remain of old revisions; older revisions had many blobs in the main repository, but modern coreboot moved almost all of them to third party submodule repositories..
Non-software blobs are permitted, so long as they are in an easily understood and/or well-documented format. For example, DDR training data is permitted (patterns used during memory controller initialization, specifically training, where the precise timings for the RAM are brute-forced); this is not software.
SPD data stored in the coreboot Git repository is in all cases some format that's simply more efficient to store as a binary, in a format that is in fact known/understood (see: coreboot source code and data sheets); in many cases, there's only one correct way to write such data, making even the question of copyright a moot point. Data is data, and code is code; the two are separate.
Non-software blobs must be redistributable under a free license, and must not be encumbered by DRM, or they will not be included in Libreboot.
Logic (in coreboot) for loading or executing binary blobs should not be removed/disabled. Libreboot merely excludes the blobs themselves. Most of the blobs that Libreboot removes (when downloading coreboot, in the build system) are CPU microcode updates; Libreboot leaves the code for loading microcode updates intact, and you can in fact insert microcode updates into your ROM image. This behaviour is intentional, and must not be removed. The only job Libreboot has is to not distribute those blobs itself!
That's all. Furthermore, Libreboot must only support systems where all of the main boot flash can be free. For example, ivybridge and sandybridge intel platforms are completely libre in coreboot, but you still need neutered Intel ME firmware in the flash, making those machines unsuitable for Libreboot.
Other firmware, such as Embedded Controller firmware, is currently outside the scope of the Libreboot project, but not due to lack of desire; rather, these are not yet possible on most supported or otherwise capable platforms, at least not with free software. Other examples of firmware outside of the main boot flash is covered in the Libreboot FAQ.
Problems with RYF criteria
You can read those guidelines by following these hyperlinks:
The FSF RYF guidelines state the following:
"However, there is one exception for secondary embedded processors. The exception applies to software delivered inside auxiliary and low-level processors and FPGAs, within which software installation is not intended after the user obtains the product. This can include, for instance, microcode inside a processor, firmware built into an I/O device, or the gate pattern of an FPGA. The software in such secondary processors does not count as product software."
This is a violation of every principle the FSF stands for, and it should be rejected on ideological grounds. The rest of libreboot's policy and overall ideology expressed, in this article, will be based largely on that rejection. The definition of product software is completely arbitrary; software is software, and software should always be free. Instead of making such exceptions, more hardware should be encouraged, with help given to provide as much freedom as possible, while providing education to users about any pitfalls they may encounter, and encourage freedom at all levels. When an organisation like the FSF makes such bold exceptions as above, it sends the wrong message, by telling people essentially to sweep these other problems under the rug, just because they involve software that happens to run on a "secondary processor". If the software is possible to update by the user, then it should be free, regardless of whether the manufacturer intended for it to be upgraded or not. Where it really isn't possible to update such software, proprietary or not, advice should be given to that effect. Education is important, and the FSF's criteria actively discourages such education; it creates a false hope that everything is great and wonderful, just because the software on one arbitrary level is all free.
This view of the FSF's, as expressed in the quoted paragraph, assumes that there is primarily one main processor controlling your system. On many modern computers, this is no longer true.
Free software does not exist in a vacuum, but we had less freedom in the past, especially when it came to hardware, so software was our primary focus.
The four freedoms are absolute, but there is a lot of nuance when it comes to boot firmware, nuance which is largely non-existent outside of firmware development, or kernel development. Most typical application/system software is high level and portable, but boot firmware has to be written for each specific machine, and due to the way hardware works, there are many trade-offs made, including by the FSF when defining what standards should apply in practise.
The fact that almost nobody talks about the EC firmware is because of the Respects Your Freedom certification. In reality, the EC firmware is crucial to user freedom, and ought to be free, but it is completely disregarded by the FSF as part of the hardware. This is wrong, and the FSF should actively actively encourage people to free it, on every laptop!
Other firmware currently outside the reach of the Libreboot project are covered in the Libreboot FAQ. For example, HDD/SSD firmware is covered in the FAQ. Again, completely disregarded and shrugged off by the FSF.
The Libreboot project will not hide or overlook these issues, because they are indeed critical, but again, currently outside the scope of what lbmk does. At the moment, lbmk concerns itself just with coreboot, but this ought to change in the future.
Examples of FSF sweeping blobs under the rug
Over the years, there have been numerous cases where the FSF actively fails to provide incentive for levels of software freedom, such as:
- TALOS II "OpenPOWER" workstation from Raptor Engineering USA. It contains a broadcom NIC inside it which requires firmware, and that firmware was non-free. The FSF were willing to ignore it, and certify the TALOS II product under RYF, but Timothy Pearson of Raptor Engineering had it freed anyway, without being told to. Hugo Landau reverse engineered the specification and Evan Lojewski wrote free firmware. See: See: https://www.devever.net/~hl/ortega and https://github.com/meklort/bcm5719-fw
- FSF once endorsed the ThinkPad X200, as sold by Minifree Ltd, which contains the Intel ME; the bootrom is still there, as is the ME coprocessor, but the ME is put into a disabled state via the Intel Flash Descriptor, and the ME firmware in flash is removed. However, the ME is an entire coprocessor which, with free firmware, could be used for a great many things. In the Libreboot and coreboot projects, there has always been interest in this but the FSF disregards it entirely. The X200 product they certified came with Libreboot pre-installed.
- Libreboot has a utility written by I, Leah Rowe, that generates ICH9M flash descriptors and GbE NVM images from scratch. This is necessary to configure the machine, but these are binary blobs otherwise; the FSF would have been quite content to certify the hardware anyway since these were non-software blobs in a format fully documented by Intel (they are just binary configuration files), but I went ahead and wrote ich9gen anyway. With ich9gen, you can more easily modify the descriptor/gbe regions for your firmware image. See: https://libreboot.org/docs/install/ich9utils.html - osboot also has this
- FSF once endorsed the ThinkPad T400 with Libreboot, as sold by Minifree. This machine comes in two versions: with ATI+Intel GPU, or only Intel GPU. If ATI GPU, it's possible to configure the machine so that either GPU is used. If the ATI GPU is to be used, a firmware blob is needed for initialization, though the driver for it is completely free. FSF ignored this fact and endorsed the hardware, so long as Libreboot does not enable the ATI GPU or tell people how to enable it. The Intel GPU on that machine has free initialization code by the coreboot project, and a fully free driver in both Linux and BSD kernels. In the configuration provided by Libreboot, the ATI GPU is completely disabled and powered down.
- All Libreboot-compatible ThinkPads contain non-free Embedded Controller firmware, which is user-flashable (and intended for update by the manufacturer). The FSF chose to ignore the EC firmware, under their secondary processor exemption. See: http://libreboot.org/faq.html#ec-embedded-controller-firmware
In all of the above cases, the FSF could have been stricter, and bolder, by insisting that these additional problems for freedom were solved. They did not. There are many other examples of this, but the purpose of this article is not to list all of them (otherwise, a book could be written on the subject).
Problems with FSDG
The FSDG criteria is separate from RYF, but has similar problems. FSDG is what the FSF-endorsed GNU+Linux distros comply with. Thoughts:
- Excluding firmware blobs in the linux kernel is bad. Non-free firmware is also bad. Including them is a wiser choice, if strong education is also provided about why they are bad (lack of freedom). If you expose them to the user, and tell them about it, there is greater incentive (by simple reminder of their existence) to reverse engineer and replace them.
- Firmware in your OS kernel is good. The FSF simultaneously gives the OK for hardware with those same firmware blobs if the firmware is embedded into a ROM/flash chip on the device, or embedded in some processor. If the firmware is on separate ROM/flash, it could still be replaced by the user via reverse engineering, but then you would probably have to do some soldering (replace the chip on the card/device). If the firmware is loaded by the OS kernel, then the firmware is exposed to the user and it can be more easily replaced. FSF criteria in this regard encourages hardware designers to hide the firmware instead, making actual freedom less likely!
Besides this, FSDG seems OK. Any free operating system should ideally not have non-free drivers or applications.
Hardware manufacturers like to shove everything into firmware because their product is often poorly designed, so they later want to provide workarounds in firmware to fix issues. In many cases, a device will already have firmware on it but require an update supplied to it by your OS kernel.
The most common example of non-free firmware in Linux is for wifi devices. This is an interesting use-case scenario, if freed, because it could be used for owner-controlled software defined radio.
The Debian way is ideal. The Debian GNU+Linux distribution is entirely free by default, and lacks any of the non-free firmware, but they have a separate repository containing non-free software. If you only want firmware, it is trivial to get installer images with it included, or add that to your installed system.
Banning linux-firmware specifically is a threat to freedom in the long term, because new users of GNU+Linux might be discouraged from using the OS if their hardware doesn't work. You might say: just buy new hardware! This is often not possible for users, and the user might not have the skill to reverse engineer it either.
More detailed insight about microcode
To be clear: it is preferable that microcode be free. The microcode on Intel and AMD systems are non-free. Facts and feelings rarely coincide; the purpose of this section is to spread facts.
Not including CPU microcode updates is an absolute disaster for system stability and security, and yet, this is one of Libreboot's key policies, to comply with FSF criteria.
Making matters worse, that very same text quoted from the FSF RYF criteria in fact specifically mentions microcode. Quoted again for posterity:
"However, there is one exception for secondary embedded processors. The exception applies to software delivered inside auxiliary and low-level processors and FPGAs, within which software installation is not intended after the user obtains the product. This can include, for instance, microcode inside a processor, firmware built into an I/O device, or the gate pattern of an FPGA. The software in such secondary processors does not count as product software."
Here, it is discussing the microcode that is burned into mask ROM on the CPU itself. It is simultaneously not giving the OK for microcode updates supplied by either coreboot or the Linux kernel; according to the FSF, these are an attack on your freedom, but the older, buggier microcode burned into ROM is OK. This is absolutely inconsistent.
The CPU already has microcode burned into mask ROM. The microcode configures logic gates in the CPU, to implement an instruction set, via special decoders which are fixed-function; it is not possible, for example, to implement a RISCV ISA on an otherwise x86 processor. It is only possible for the microcode to implement x86, or broken x86, and the default microcode is almost always broken x86 on Intel/AMD CPUs; it is inevitable, due to the complexity of these processors.
The basis of the FSF's disagreement about microcode updates is that they do believe otherwise; Stallman himself expressed such ignorance to me, in a recent email conversation I had with him, as of January 2nd, 2022. The FSF believes that these x86 microcode updates (on Intel/AMD) allow you to completely create a new CPU that is fundamentally different than x86. This is not true. It is also not true that all instructions in x86 ISA are implemented with microcode. In some cases, hardcoded circuitry is used! The microcode updates are more like tiny one liner patches here and there in a git repository, by way of analogy. To once again get in the head-space of the FSF: these updates cannot do the CPU equivalent of re-factoring an entire codebase. They are hot fixes, nothing more!
These processors provide a way to supply microcode updates. These updates
are volatile, and consequently must be applied during every boot cycle. The
updates fix stability/reliability/security bugs, and their absence
is technically incorrect, but Libreboot excludes them anyway, because that is
FSF policy. Examples of where these updates fix bugs: on ASUS KCMA-D8/KGPE-D16
and ThinkPad X200/T400/T500/W500/X200T/X200/R500/X301, the updates make
hardware-based virtualization (via kvm) completely stable, where it would
otherwise lead to a kernel panic. They allow those same thinkpads to be run with
high CPU usage and I/O (RAM usage), without crashing (otherwise, it's very
likely to encounter a kernel panic caused by a
Machine Check Exception).
Not including these updates will result in an unstable/undefined state. Intel themselves define which bugs affect which CPUs, and they define workarounds, or provide fixes in microcode. Based on this, software such as the Linux kernel can work around those bugs/quirks. Also, upstream versions of the Linux kernel can update the microcode at boot time (however, it is recommend still to do it from coreboot, for more stable memory controller initialization or “raminit”). Similar can be said about AMD CPUs.
Here are some examples of where lack of microcode updates affected Libreboot, forcing Libreboot to work around changes made upstream in coreboot, changes that were good and made coreboot behave in a more standards-compliant manner as per Intel specifications. Libreboot had to break coreboot to retain certain other functionalities, on some GM45/ICH9M thinkpads:
These patches revert bug fixes in coreboot, fixes that happen to break other functionality but only when microcode updates are excluded. The most technically correct solution is to not apply the above patches, and instead supply microcode updates!
Pick your poison. Libreboot does not disable the mechanism in coreboot to load these updates. At boot time, coreboot can supply such updates to the CPU, if present in CBFS. Libreboot merely excludes them, but you can add them to your Libreboot ROM image. A fork of Libreboot, named osboot, includes them by default; it does this, even on libreboot-compatible hardware. Not adding the updates is irresponsible, but a promise was made to the FSF back in 2013 when the Libreboot project started, precisely that it would not add microcode to ROM images by default. It is Libreboot's policy to keep that promise, despite the obvious flaw of that policy.
More info about osboot is available on https://osboot.org/ - osboot's policy is the same as Libreboot, except that it does not delete blobs; the goal is still software freedom, but it provides those users who are not willing/able to use libreboot hardware to otherwise still have some freedoms compared to otherwise fully proprietary vendor firmware. osboot and libreboot are two sides of a coin; libreboot is the "light", and osboot is the dark side. Both projects are maintained and were founded by Leah Rowe.
The osboot firmware is far superior to Libreboot, in terms of reliability, due to the presence of microcode updates in the firmware, and with zero practical change to your freedom, on libreboot-compatible hardware. However, I will say:
People have been using Libreboot for years, on these machines, and most people don't really have that many issues, most of the time. My opposition to FSF's microcode policy is out of principle. Logical, common sense principle. I simply cannot compute that microcode updates are an attack on your freedom, because:
Microcode updates are not an attack on your freedom. The FSF's opposition to these updates is both symbolic and ignorant; it is ultimately futile, but I digress.
I will continue to develop Libreboot and osboot, in parallel.
Other considerations
Also not covered strictly by Libreboot: OSHW and Right To Repair. Freedom at the silicon level would however be amazing, and efforts already exist; for example, look at the RISCV ISA (in practise, actual fabrication is still proprietary and not under your control, but RISCV is a completely free CPU design that companies can use, instead of having to use proprietary ARM/x86 and so on). Similarly, Right To Repair (ability to repair your own device, which implies free access to schematics and diagrams) is critical, for the same reason that Free Software (Right To Hack) is critical!
OSHW and Right To Repair are not covered at all by RYF (FSF's Respects Your Freedom criteria), the criteria which Libreboot was created to comply with. RYF also makes several concessions that are ultimately damaging, such as the software as circuitry policy which is, frankly, nonsensical. ROM is still software. There was a time when the FSF didn't consider BIOS software a freedom issue, just because it was burned onto a mask ROM instead of flashed; those FSF policies ignore the fact that, with adequate soldering skills, it is trivial to replace stand-alone mask ROM ICs with compatible flash memory.
Conclusion
Compromise and nuance is the name of the game, even if you're the FSF. It is completely unavoidable, but there are some who try to deny this fact and pretend like things are as they'd prefer them to be, rather than how they actually are in the real world.
Facts and feelings are usually very different things, and contradictory.
RYF isn't wrong per se, just flawed. It is correct in some ways and if complied with, the result does give many freedoms to the user, but RYF completely disregards many things that are now possible, including freedoms at the hardware level (the RYF criteria only covers software). Those guidelines are written with assumptions that were still true in the 1990s, but the world has since evolved. As of 2 January 2022, Libreboot still complies strictly with RYF, and will continue to do so, at least for the time being.
The conclusion that should be drawn from all of this is as follows:
Following FSF criteria does not damage anything, but that criteria is very conservative. Its exemptions should be disregarded and entirely ignored. RYF is no longer fit for purpose, and should be rewritten to create a more strict set of guidelines, without all the loopholes or exemptions. As has always been the case, Libreboot tries to always go above and beyond, but the Libreboot project does not see RYF as a gold standard. There are levels of freedom possible now that the RYF guidelines do not cover at all, and in some cases even actively discourage/dis-incentivize because it makes compromises based on assumptions that are no longer true.
Sad truth: RYF actively encourages less freedom, by not being bold enough. It sets a victory flag and says mission accomplished, despite the fact that the work is far from complete!
If followed with exemptions unchallenged, RYF may in some cases encourage companies to sweep under the rug any freedom issues that exist, where it concerns non-free firmware not running on the host CPU (such as the Embedded Controller firmware).
I propose that new guidelines be written, to replace RYF. These new guidelines will do away with all exemptions/loopholes, and demand that all software be free on the machine, or as much as possible. Instead of only promoting products that meet some arbitrary standard, simply catalog all systems on a grand database of sorts (like h-node.org, but better). Include Right to Repair and OSHW (including things like RISCV) in the most "ideal" standard machine.
Don't call it "Respects Your Freedom" or something similar. Instead, call it something like: the freedom catalog. And actually focus on hardware, not just software!
In the year 2022 onwards, we can do better. The RYF program should be cancelled. It is no longer fit for purpose.
Other resources
Ariadne Conill's RYF blog post
Ariadne Conill, security team chair of Alpine Linux, posted a very robust article about RYF, with similar points made when compared to this article. However, Ariadne goes into detail on several other examples of problems with the FSF RYF criteria; for example, it talks about the Novena product by Bunnie.
It's worth a read! Link:
Hector Martin's RYF thread
Hector Martin, leader of the Asahi Linux project (for booting linux kernels on M1 macbooks) wrote a very robust twitter thread criticizing the RYF criteria and much of what he wrote inspired this article that you are reading. See:
https://twitter.com/marcan42/status/1040626210999431168
If you wish to avoid non-free javascript, you can read this thread using nitter:
https://nitter.net/marcan42/status/1040626210999431168
Article updates
23 January 2022
Added link to Ariadne Conill's article.
21 January 2022
This article was updated on 21 January 2022, to add the section with examples in the real world of FSF sweeping blobs under the rug (ATI T400 thinkpads, ICH9M descriptors and TALOS II NIC firmware).
Also on 21 January 2022: added section about FSDG (criticisms of it).
Also on 21 January 2022: added link to Hector Martin's twitter thread.